← Home

REVIEWED & READY TO INSTALL

BLA (Blind LLM Agents)

qybaihe/bla

Open-source CLI and MCP server that protects your own software against AI reverse-engineering agents such as REA: it plants decoy clues, never-executed logic and analyzer budget traps so the agent misreads or gives up, while normal users see no difference. Free, MIT license.

STEP BY STEP

How to install BLA (Blind LLM Agents)

The step-by-step guide is right on this page, written for people who have never installed anything from GitHub.

Updated 11/14/2026
GitHub preview image of BLA (Blind LLM Agents)
GitHub stars
6,383
Language
TypeScript
Added
11/14/2026

How to install BLA (Blind LLM Agents), step by step

Work from top to bottom. Each grey box is one command: copy the whole line, paste it into the terminal window and press Enter.

Never used a terminal? Read this first (1 minute)

A terminal is a window where you type commands for your computer. You do not need to understand them, just copy and paste them exactly.

  1. Windows: press the Windows key, type "PowerShell", then press Enter.
  2. macOS: press Command + Space, type "Terminal", then press Enter.
  3. Linux: press Ctrl + Alt + T.
  4. Copy a command from the guide, right-click inside the window you just opened to paste it (macOS: Command + V), then press Enter.
  5. Wait for the command to finish (the cursor blinks again on a new line) before running the next one.

Lines starting with # are notes; you do not need to copy them. If red error text appears, check the common errors section at the end of the guide.

1. What is BLA?

BLA (short for Blind LLM Agents; repo qybaihe/bla, npm package bla-agents) is an open-source command-line tool that protects your own software against AI reverse-engineering agents such as REA. Reverse engineering means reading a packaged program backwards to rebuild its source code. Instead of classic obfuscation, BLA plants fake clues, decoy logic that never runs, and structures that deliberately exceed the limits of analysis tools, so an AI agent draws the wrong conclusions or gives up. Your users run the program as usual.

BLA is free under the MIT license. Besides the CLI it ships an MCP server, which lets AI coding assistants such as Claude Code, Cursor or Codex call BLA's tools directly.

To be clear: this is a tool for people who build and ship software. If you only use apps, you do not need it.

A good fit if you:

  • Build and release software: built JavaScript bundles, Electron apps, WebAssembly, Python, native executables, SQLite databases, GGUF/Safetensors model files, and more.
  • Worry that important logic (a license check, for example) could be cloned quickly and cheaply by an AI agent.
  • Are comfortable running npx, or already use an AI coding assistant.

Not a good fit if you:

  • Want to dissect someone else's software. The official docs restrict BLA to software you own or are authorized to modify and distribute.
  • Need absolute secrecy. The docs admit no client-side scheme can stop a dedicated human analyst with time and debugging tools. Keep high-value keys on your own servers.
  • Cannot accept larger builds. The README's results table shows some files growing from a few hundred bytes to several megabytes.

About the numbers: the README cites "300× to 2,800×" more tokens for an attacker. That is the project's own modeled estimate and has not been independently verified.

2. What you need

  • A computer running Windows, macOS or Linux.
  • Node.js 22.23.4 or later on the 22 line, or 24.0.0 or later (the package's engines field). Node.js is the runtime for JavaScript programs and includes the npm and npx commands. Free download: https://nodejs.org (pick a current LTS).
  • Your project folder, already built into the files you ship (usually dist/, build/, out/ or release/).
  • A backup of the project, or the project under Git. BLA writes to a separate folder by default, but a backup is still a good habit.
  • (Optional) An AI coding assistant if you prefer to give instructions in plain language: Claude Code, Cursor, Codex, Windsurf, Gemini CLI or OpenCode.

No account or API key is needed for BLA itself. If you drive it through an assistant, that assistant's own pricing applies.

3. Step-by-step installation

BLA has no installer. The easiest path is to run it directly with npx, with nothing installed permanently.

Step 1: Check Node.js

Open Terminal (macOS/Linux) or PowerShell (Windows) and type:

node --version

What you should see: a line such as v24.x.x, or v22.23.4 or later. If you get "command not found" or a lower version, install the current Node.js LTS from https://nodejs.org and reopen the terminal.

Step 2: Go to your project folder

Use cd to enter the root of your project (the folder with package.json or a dist/ folder):

cd path-to-your-project

What you should see: the prompt shows your project folder. ls (macOS/Linux) or dir (Windows) lists your project files.

Step 3: Preview what BLA will touch

Before hardening anything, run the preview mode. It writes zero bytes to disk:

npx bla-agents harden . --dry-run

The first time, npx asks whether to download bla-agents; type y and press Enter.

What you should see: a list of files BLA found to protect (files in dist/ or build/, .asar, .wasm, .sqlite and so on). No new folder is created. According to the docs, a dry run never reports overall success, so an "incomplete" result here is expected.

Step 4: Harden for real

npx bla-agents harden .

This discovers your build outputs and runs three stages: audit (exposure check), protect (inject the shields) and verify.

What you should see: an Output: line near the end pointing to a .bla-hardened folder inside your project. The hardened copies live there under their original names; your original files are not overwritten.

Note: BLA refuses to overwrite an existing file without confirmation. If an output already exists it stops and lists what would be overwritten. Read that list before agreeing.

Step 5 (optional): Register BLA with your AI coding assistant

npx bla-agents setup

To register only some assistants:

npx bla-agents setup --client claude-code,cursor

Then check the result:

npx bla-agents doctor

What you should see: setup reports that the bla MCP server and the blind-llm-agents skill were registered for the assistants it found. doctor prints the readiness of the protection engines and of each registered assistant.

Step 6 (optional): Install permanently

To harden automatically as part of your build, add BLA as a dev dependency:

npm install --save-dev bla-agents

Or install it globally to get the short bla command:

npm install -g bla-agents

What you should see: bla --version prints a version number (3.3.1 was the latest when this guide was written).

4. First use

The quickest way to see results is to harden one file and verify it:

  1. Check how exposed your file is before release:
npx bla-agents audit ./dist/app.js
  1. Protect it at the strongest level (bankrupt):
npx bla-agents protect ./dist/app.js --mode bankrupt
  1. Compare the protected copy with the original:
npx bla-agents verify ./dist/app.js ./dist/app.bla.js

Replace ./dist/app.js with your real file. verify checks the counter-REA structures and estimates how many extra tokens an attacker would burn.

Important: the official docs ask you to run your own project's test suite against the protected build in a staging environment before shipping. verify checks counter-REA structure only; it does not replace testing your real app.

If you completed Step 5, open your assistant in the project folder and type: Harden my software against REA. The assistant calls BLA's tools to run the steps above.

5. Common errors and fixes

SymptomCauseFix
npx: command not found or node does not runNode.js missing, or the terminal was opened before installingInstall Node.js LTS from https://nodejs.org, then close and reopen the terminal
Unsupported Node versionNode older than 22.23.4Upgrade to Node 24 LTS or the latest 22
harden finds no filesProject not built yet, or wrong folderRun your build first, then cd into the project root
Stops with "confirmation required"An output file already exists and BLA will not overwrite it without permissionChoose another output folder, or delete the old .bla-hardened folder after checking it
Refuses bankrupt because of plaintext secretsThe bla lint gate found strings like FLAG{...}, keys, or a single easily patched checkRun npx bla-agents lint . to see the findings, fix them in source and rerun
Hardened app misbehavesAn edge case in your projectGo back to the original (BLA does not overwrite it) and report it on the repo's Issues page
setup finds no assistantAssistant not installed, or its config lives elsewhereInstall the assistant first or pass --client, then run npx bla-agents doctor

6. Uninstall / update

Remove from your AI assistants (MCP registration and skill folders):

npx bla-agents uninstall

Remove a global install:

npm uninstall -g bla-agents

Remove from a project:

npm uninstall bla-agents

You can delete the .bla-hardened folder created by harden by hand when you no longer need it.

Update: with npx, add @latest to always run the newest release, for example npx bla-agents@latest harden .. For project or global installs, rerun the matching npm install. The project releases very often; see https://github.com/qybaihe/bla/releases

7. FAQ

Does BLA cost money? No. It is open source under the MIT license.

Will my app get slower? According to the docs, decoy branches sit behind mathematical conditions that are always false, so they never run during normal use, and the project's own test targets produce the same output as the originals. Run your own tests to be sure. Files do get larger.

Does BLA inject malware or send data anywhere? The docs state BLA never injects destructive commands, shell payloads or network callbacks into protected files. The source is public so you can check.

Does it stop manual reverse engineering? Not fully. BLA targets automated AI reverse-engineering agents. The docs make no promise against an expert using tools such as Ghidra or Frida over time.

Can I use BLA on someone else's software? No. Only on software you own or are authorized to modify and distribute.

Written by GitHot and checked against the project's official documentation. If a step is wrong or unclear, message GitHot through the channels at the bottom of the page.

SEE WHY IT'S HOT

Watch the original review

Go back to the video that brought you here.

Watch on TikTok ↗

Don't miss the next repo

Every new repo gets a video review, with its install guide right here.

Follow on TikTok ↗